nd-skills

ND Skills — 25 agent skills: 24 focused behaviors plus one manual router.

Diagnosis-free neurodivergent-support behaviors for coding agents: executive-function routing, functional minimums, task decomposition, environmental scaffolding, emotion regulation, hyperfocus exits, task boundaries, working-memory support, realistic time estimates, and more.

Works with any agent that reads Markdown skills, including Claude Code and Codex.

What's New

v2.2.0

Companion pack: autism-layer (literal reading, scripts, invitations), biz-weekly-review (coach-style week review), nd-interrupt-budget (silent by default, a nudge budget the person sets), nd-inbox-digest (at most three literal actions, read-only).
Ambient signals section: what a host must supply for turn-count triggers, and the working-memory schema gap.

v2.1.0

Adds the manual nd router for one-command access to the full pack.

v2.0.0

Six new skills for routing, functional minimums, decomposition, goal persistence, environmental scaffolding, and emotion regulation.
Conversation-derived fixes for constraint loss, premature stopping, false completion, and unnecessary clarification questions.
Synchronized always-on instructions for Claude Code and Codex while preserving surrounding configuration.
Redacted conversation-audit tooling and deterministic behavior-regression cases.

Install

For invocable skills only:

npx skills add jpoindexter/nd-skills

For invocable skills plus synchronized always-on behavior in ~/.claude/CLAUDE.md and ~/.codex/AGENTS.md:

git clone https://github.com/jpoindexter/nd-skills.git
cd nd-skills
./install.sh

The installer updates its managed blocks on every run, preserves surrounding content, and copies complete skill folders including agents/openai.yaml metadata.

Restart Claude Code or Codex after installation so the refreshed skill inventory and always-on instructions are loaded.

Always On and Manual Invocation

The installer makes the 24 focused behaviors always-on by synchronizing a managed block into Claude Code and Codex. The block is loaded at the start of new sessions and routes only from explicit task signals; it does not load every full skill body into every response.

Use the nd umbrella skill when you want to invoke the router yourself:

Claude Code: /nd [task]
Codex:       $nd [task]

Claude Code resolves installed skills directly as slash commands. Current Codex versions use skill mentions or the skills picker; custom slash prompts were removed, so Codex does not support a literal custom /nd. Use $nd there.

You can also use /nd status or $nd status to show the active constraint ledger, and off to remove optional ND scaffolding for the current task.

How It Works

Invocable layer: every folder under skills/ is a standalone agent skill with trigger metadata and focused instructions.
Always-on layer: install.sh synchronizes a managed instruction block into Claude Code and Codex without replacing unrelated user configuration.
Behavior layer: the router selects the smallest useful support, while constraint and goal state stay explicit across longer tasks.
Evaluation layer: synthetic regression cases preserve the rules derived from real conversation failures without committing private conversation archives.

The Skills

Skill	Covers
`nd`	Manually routes a task through the smallest relevant subset of the pack. Invoke as `/nd` in Claude Code or `$nd` in Codex.
`nd-executive-function-router`	Routes explicit task-state barriers to the least intrusive support.
`nd-functional-minimums`	Preserves safe, useful outcomes when capacity is low.
`nd-task-decomposition`	Converts oversized work into verifiable Now, Next, Later, Re-entry, and Done units with restart-safe authority refresh.
`nd-goal-persistence`	Carries terminal conditions across subtasks until end-to-end evidence or a genuine blocker.
`nd-environment-scaffold`	Changes cues, defaults, tools, and support at the point of action.
`nd-emotion-regulation`	Protects work and decision quality during explicit emotional intensity.
`nd-choicereduce`	Shows only the top three options when a larger choice set would add load.
`nd-closuregate`	Surfaces nearly finished work before a substantially new task begins.
`nd-complexity-gate`	Suggests structured planning before complex or ambiguous execution.
`nd-hyperfocus-guard`	Detects deep rabbit holes and offers a route back to the original goal.
`nd-inhibit`	Checks significant actions against the active goal and names repeated drift.
`nd-research-gate`	Interrupts research spirals that produce no concrete output.
`nd-selfmonitor`	Adds a visible intent check before destructive or hard-to-reverse actions.
`nd-sensory-load`	Makes communication compressed, predictable, layered, and controllable.
`nd-setshift`	Detects repeated failed approaches and forces a genuinely different strategy.
`nd-task-initiation`	Prescribes a concrete first action and a truthful activation bridge when needed.
`nd-taskboundary`	Saves restartable state and marks clean context transitions.
`nd-time-blindness`	Externalizes ranges, hidden costs, checkpoints, and stopping rules.
`nd-velocity-check`	Makes the ideas-captured to items-shipped ratio visible.
`nd-working-memory`	Re-anchors active context and persists one durable working surface.
`autism-layer`	Reads messages literally in four labelled lines, drafts scripts in the person's voice, phrases every suggestion as an invitation, protects special interests, names masking plainly.
`biz-weekly-review`	Sorts the week into motion vs action, picks one front, gives at most three commitments and one hard recommendation.
`nd-interrupt-budget`	Silent by default; unprompted nudges spend a daily budget the person set, one sentence each with a Not now, never during quiet mode.
`nd-inbox-digest`	Turns an inbox into at most three literal one-line actions, read-only, never sends, lists what the model saw.

Ambient signals

Several skills fire on counts: nd-research-gate after 8 research turns, nd-hyperfocus-guard after 10 turns in one area, nd-working-memory every 5 turns. A turn count is the only signal a chat agent has, and it is a weak proxy for what those skills are about (time lost, attention gone elsewhere). A host that supplies three ambient signals can fire them on the real thing:

idle time — seconds since the last keyboard or mouse input
front app — which application is in front, by name or bundle id, not its window title or URL
clock — local time, so a rule can say "20 minutes" instead of "8 turns"

Example: Heel, a macOS companion, reads the front app and idle seconds every 5 s with no permission prompt, keeps per-app seconds for the day only, never sends them to the model, and turns them into one-sentence nudges inside nd-interrupt-budget. A host without these signals keeps the turn-count triggers as they are.

Gap: nd-working-memory names one durable working surface but this repo gives it no schema. Proposed shape, one record per task: outcome, target, must, must_not, authority, done_evidence. Not added here; a host that persists it should say where.

Source Basis

The pack synthesizes and paraphrases task-support concepts from:

Peg Dawson and Richard Guare, the Smart but Scattered series
Judith Kolberg and Kathleen Nadeau, ADD-Friendly Ways to Organize Your Life
KC Davis, How to Keep House While Drowning
Tamara Rosier, Your Brain's Not Broken Workbook
Megan Anna Neff, Self-Care for Autistic People
Devon Price, Unmasking Autism
University of Oxford, NESTL Toolkit

No source text, diagnostic assessment, or personal profile is bundled. The skills modify tasks, environments, and agent behavior; they are not medical advice.

Validation

Run the deterministic behavior checks:

python3 scripts/check_behavior_regressions.py
bash -n install.sh

The sanitized findings and limitations are documented in the Claude + Codex conversation audit.

License

MIT — see LICENSE.